diff --git a/pymongo/change_stream.py b/pymongo/change_stream.py
index ad3dbaa61..8b2b57289 100644
--- a/pymongo/change_stream.py
+++ b/pymongo/change_stream.py
@@ -66,9 +66,11 @@ class ChangeStream(object):
         self._orig_codec_options = target.codec_options
         if target.codec_options.type_registry._decoder_map:
             self._decode_custom = True
+            # Keep the type registry so that we support encoding custom types
+            # in the pipeline.
             self._target = target.with_options(
                 codec_options=target.codec_options.with_options(
-                    document_class=RawBSONDocument, type_registry=None))
+                    document_class=RawBSONDocument))
         else:
             self._target = target
 
diff --git a/test/test_custom_types.py b/test/test_custom_types.py
index cc83c1e6d..d7095da38 100644
--- a/test/test_custom_types.py
+++ b/test/test_custom_types.py
@@ -777,6 +777,29 @@ class ChangeStreamsWCustomTypesTestMixin(object):
         self.kill_change_stream_cursor(change_stream)
         self.insert_and_check(change_stream, input_docs[2], expected_docs[2])
 
+    def test_custom_type_in_pipeline(self):
+        codecopts = CodecOptions(type_registry=TypeRegistry([
+            UndecipherableIntEncoder(), UppercaseTextDecoder()]))
+        self.create_targets(codec_options=codecopts)
+
+        input_docs = [
+            {'_id': UndecipherableInt64Type(1), 'data': 'hello'},
+            {'_id': 2, 'data': 'world'},
+            {'_id': UndecipherableInt64Type(3), 'data': '!'}]
+        expected_docs = [
+            {'_id': 2, 'data': 'WORLD'},
+            {'_id': 3, 'data': '!'}]
+
+        # UndecipherableInt64Type should be encoded with the TypeRegistry.
+        change_stream = self.change_stream(
+            [{'$match': {'documentKey._id': {
+                '$gte': UndecipherableInt64Type(2)}}}])
+
+        self.input_target.insert_one(input_docs[0])
+        self.insert_and_check(change_stream, input_docs[1], expected_docs[0])
+        self.kill_change_stream_cursor(change_stream)
+        self.insert_and_check(change_stream, input_docs[2], expected_docs[1])
+
     def test_break_resume_token(self):
         # Get one document from a change stream to determine resumeToken type.
         self.create_targets()